Research On Salient Object Detection Based On Convolutional Neural Network

Salient object detection aims to highlight visually significant areas in image and is applied to various computer vision tasks.However,there are still many problems in salient object detection,such as multi-scale,foreground misjudgment,background misjudgment and boundary complexity.In order to solve these problems effectively,this paper researches salient object detection based on convolutional network and deep learning,and designs end-to-end full convolutional networks to improve the detection performance.The research content includes the following three parts.1.Salient object detection based on coordinate attention feature pyramid.Compared with traditional methods,the algorithms based on full convolutional network have shown great advantages.However,full convolutional network still fails to solve the problem of background misjudgment and boundary complex.This part proposes a coordinate attention feature pyramid network to solve the two challenges.The model is based on feature pyramid to extract multilevel salient features,and designs a feature refinement module to aggregate these features.To solve the problem of background misjudgment,the model utilizes the last coordinate attention mechanism to assign different weights to deep salient features.For boundary complex problems,boundary awareness loss is proposed to monitor the learning of salient contour and help network to predict boundary pixels more accurately.Experimental results on five public salient object detection datasets show that the proposed method can effectively improve the performance of detection.2.Salient object detection based on dual decoders and boundary awareness.Although the methods based on full convolutional network are widely used in salient object detection,the problems of foreground misjudgment,background misjudgment and complex boundary still exist in complex scenes.The dual decoders and boundary awareness network is designed to solve these problems.In order to effectively deal with foreground background misjudgment,the model propose a dual decoders structure of foreground decoding branch and background decoding branch,in which foreground decoding branch aims to learn salient region features to enhance foreground information,while background decoding branch aims to learn non-salient region information and suppress background noise.For boundary complexity problems,the model does not design boundary awareness loss function,but proposes a boundary awareness module.The module uses boundary maps to help the module learn the contour information of salient objects,so as to generate clear boundaries.The proposed method is validated on five commonly salient object detection datasets,and the results show that the method has greater advantages than the current advanced methods.3.Salient object detection based on multi-scale feature aggregation and boundary awareness.For the salient object detection based on convolutional neural network,there are multi-scale and boundary complexity problems.This part propose a salient object detection model based on multi-scale feature aggregation and boundary awareness.A multi-scale feature aggregation module is proposed,and a multi-level feature aggregation strategy is adopted to learn the features of objects with different scales.At the same time,a cross feature refinement module is proposed to further integrate multi-scale features in a more effective way.For complex boundary problems,the algorithm redesigned boundary pixels awareness loss,making it pay more attention to the pixels closer to the boundary,so as to generate clear and accurate boundaries.Sufficient experiments show that the performance of this method is better than most advanced salient object detection methods.